AI pentest methodology in 8 steps
In short
An AI pentest follows the same phases as a traditional penetration test (scoping, reconnaissance, analysis, exploitation, reporting, debrief), but enumeration, analysis and drafting are accelerated by AI agents. Decision phases (scoping, sensitive exploitation, validation, debrief) remain human. HackHeart's methodology is based on the OWASP WSTG, PTES and MITRE ATT&CK.
The 8 steps of an AI pentest at HackHeart
- Scoping and authorisation. Targets, testing windows, provided accounts and exclusions are defined and a written authorisation is signed. The scope is then hard-coded into the agents' configuration.
- Preparation. Test accounts per role, black, grey or white box choice, pseudonymisation set-up.
- Assisted reconnaissance. Agents map subdomains, exposed services, endpoints, parameters and technologies. The tester validates and prioritises.
- Automated analysis. Specialised agents test the OWASP Top 10 and OWASP API Security Top 10 families.
- Controlled exploitation. Attack chains are built and exploited by the tester with agent support, without destructive actions. Internally, the same logic applies to Active Directory (Kerberoasting, delegation, ACLs, ADCS).
- Qualification. An adversarial agent tries to invalidate each finding, then the tester manually reproduces the survivors. Each vulnerability gets a CVSS score and a context-aware severity.
- Reporting. Executive summary, technical detail per vulnerability (evidence, reproduction, impact, prioritised remediation) and an action plan.
- Debrief and retest. Results are presented live, then fixes are verified after remediation.
Who does what: AI or human?
| Phase | AI agents | Penetration tester |
|---|---|---|
| Scoping | None | Full |
| Reconnaissance | Most of the work | Validation, prioritisation |
| Analysis | Most of the work | Steering, business logic |
| Exploitation | Assistance | Decision and execution |
| Qualification | First adversarial filter | Reproduction and final validation |
| Reporting | First draft | Review, context, sign-off |
| Debrief | None | Full |
Scopes covered
Web applications
Websites, extranets, customer portals, SaaS, grey box with per-role accounts.
APIs
REST and GraphQL, with a focus on object-level access control.
External perimeter
Everything an attacker sees from the Internet: services, VPN, mail, DNS.
Internal and Active Directory
Simulating a compromised workstation and finding paths to privileged accounts.